AI Writing Styles: A Content Writer’s Guide
- Large language models, like humans, exhibit distinctive writing styles, according to a new study.
- Mingjie Sun, a PhD student in computer science, expressed surprise at the 97% accuracy rate achieved in the study. The team's classifier program outperformed expectations, distinguishing between five...
- The computer analysis highlighted individual profiles for each large language model.
A groundbreaking Carnegie Mellon study reveals that AI large language models (LLMs) possess distinct writing styles,impacting the way we understand and use AI. Researchers achieved 97% accuracy in identifying the source AI model behind a text, highlighting characteristic differences. Such as, ChatGPT leans toward detailed explanations, while Claude favors concise answers. This has implications for AI training and the use of synthetic data, with potential biases transferring between models. Learn how these stylistic nuances can affect synthetic data and AI training. Explore how News Directory 3 is following the latest developments. Discover what’s next for LLMs.
AI Models Show Distinct Writing Styles, Carnegie Mellon Study Finds
Updated April 29, 2025
Large language models, like humans, exhibit distinctive writing styles, according to a new study. Researchers at Carnegie Mellon University discovered that they could pinpoint which AI model generated a text with remarkable accuracy by analyzing characteristic word choices.
Mingjie Sun, a PhD student in computer science, expressed surprise at the 97% accuracy rate achieved in the study. The team’s classifier program outperformed expectations, distinguishing between five LLMs—ChatGPT, Claude, Gemini, Grok, and Deepseek—despite the complexity of the task.
The computer analysis highlighted individual profiles for each large language model. Such as, ChatGPT often provides detailed, explanatory content, while Claude prefers delivering succinct, direct responses. These stylistic differences are deeply ingrained within each model, remaining consistent even when texts are altered through scrambling, rephrasing, translation, or summarization.
The findings raise concerns about using synthetic data, or text generated by LLMs, to train new AI models. This practice could inadvertently transfer stylistic quirks from the source model, potentially leading to unforeseen consequences in the behavior of subsequent AI generations. Zico Kolter, a professor and director at CMU, noted that the use of synthetic data for training, once common, is now declining.
The research team, including members from the University of California, berkeley; the University of Pennsylvania; and Princeton university, focused on gaining a deeper understanding of large language models rather than differentiating between AI- and human-generated text. Kolter emphasized the value of understanding the distinguishing characteristics of various models, especially given the increasing volume of LLM-generated content online. This understanding is key to responsible AI training and mitigating potential biases.
The study underscores the importance of understanding the unique characteristics of different AI models, notably as they increasingly contribute to online content. Further research into large language model behavior is essential for responsible development and deployment.
“This work is much more about understanding the distinctive characteristics, the natures of different LLMs, the same way we think about different styles of writing by peopel,” Kolter said.
What’s next
Future research may explore how these distinctive styles impact user perception and trust in AI-generated content, and also strategies for mitigating potential biases transferred through synthetic data.
